Build a taste corpus your agent can read

You can now flag anything on a page as good or bad, group those judgments into collections, and let your agent pull the corpus over MCP. The groundwork for agents that learn your preferences.

Every team has a shared sense of taste that nobody can point at. You know the spacing that feels right, the type hierarchy that reads well, the checkout flow that respects the user — and the dozen small things on the current marketing site that quietly undercut all of it. That knowledge lives in heads, in Slack threads, in the one senior designer’s memory. An AI coding agent working on your site has none of it.

HAVI’s answer is a taste corpus: a place where “I like this” and “this is wrong” are captured with the same care we already put into bug reports, so your agent can read them as data instead of never seeing them at all.

How it works

The corpus is built from taste items. A taste item is an annotation — the same W3C Web Annotation format every HAVI capture uses — captured with the assessing motivation, where feedback captures stay describing. That distinction travels with the item everywhere: the dashboard, the REST API, and MCP.

Every taste item carries:

  • A judgment — good or bad. The one bit of information that makes taste usable: not “here is a screenshot of a button,” but “this button is right” or “this button is wrong.”
  • A collection. Taste items are grouped into named collections — “Landing page polish,” “Mobile spacing,” “Competitor patterns worth stealing” — so the corpus browses like a curated set, not a stream. Every taste item belongs to exactly one collection.
  • Optionally, a project. A taste item can reference a Project, so the corpus can be scoped to “everything for the relaunch” alongside the git context HAVI already captures.

In the hosted workspace dashboard, the Taste corpus view lists your collections and their items. You create collections there, rename them, and delete them — with one deliberate guard: a collection that still holds taste items refuses to delete until you’ve moved them elsewhere. No silent orphans.

The judgment is editable after the fact, too. Change your mind, flip good to bad, or clear the judgment entirely — from the dashboard or from the capture itself.

Capturing taste where you find it

Taste happens while you’re browsing, so capture happens there too. In the browser extension’s capture composer there’s a mode toggle: feedback, as before, or taste. In taste mode you get good/bad judgment chips, a collection picker, and an optional project picker. Pick the element, judge it, file it, done — the same one-shortcut flow, now with a verdict attached.

The judgment is stored as a structured tagging body on the annotation (x:labelKey: "judgment"), not as prose in a comment. That matters for the next paragraph.

Native mobile clients capture on the same contract: taste items posted from a mobile app look exactly like taste items posted from the extension, and land in the same collections.

What your agent sees

Because taste items are ordinary annotations, the MCP read path already knows how to find them. The list_annotations tool accepts two new filters: motivation=assessing pulls the whole corpus, and collection_id pulls a single collection. A session can ask for “everything this team has flagged bad in the landing-page collection” and get back structured annotations — screenshots, selectors, judgments — without parsing a word of prose.

As always with HAVI, nothing is pushed into a running session. The agent pulls the corpus when you ask it to, and what it does next is its call: fix the code directly, enrich and file it, or something else entirely.

What this is, and isn’t yet

The honest framing, because it matters: this release is the collection layer. HAVI records your judgments and hands them to your agent as data. It does not score pages, rank items, or decide what “good taste” is — the judgment stays yours.

The natural next step is distillation: an agent reading the corpus, seeing the pattern in what a team consistently flags good and bad, and deriving preferences it can apply to future work. That step is groundwork today — the filters exist so it can be built — not a shipped feature. We’d rather ship the capture loop that fills the corpus first, because a preference engine with an empty corpus isn’t worth much.

Try it

If you already use HAVI, the taste corpus is in your workspace now: open the Taste corpus view, create a collection, and flip the extension’s capture composer to taste mode on your next walkthrough.

Try HAVI free →

Give your agent eyes on your browser.
Free tier, no card. Install the Chrome extension in under a minute.
Add to Chrome — Free